Enhancing Korean Dependency Parsing with Morphosyntactic Features

UniDive for Korean: combining Universal Dependencies with UniMorph-derived features improves dependency parsing

arXiv 2025

Jungyeul Park1, Yige Chen2, Kyuwon Kim3, KyungTae Lim4, Chulwoo Park5

1The University of British Columbia, Canada 2The Chinese University of Hong Kong, Hong Kong 3Seoul National University, South Korea 4KAIST, South Korea 5Anyang University, South Korea

Paper (arXiv) PDF Code

Abstract

This paper introduces UniDive for Korean, an integrated framework that bridges Universal Dependencies (UD) and Universal Morphology (UniMorph) to enhance the representation and processing of Korean morphosyntax. Korean's rich inflectional morphology and flexible word order pose challenges for existing frameworks, which often treat morphology and syntax separately, leading to inconsistencies in linguistic analysis. UniDive unifies syntactic and morphological annotations by preserving syntactic dependencies while incorporating UniMorph-derived features, improving consistency in annotation. We construct an integrated dataset and apply it to dependency parsing, demonstrating that enriched morphosyntactic features enhance parsing accuracy, particularly in distinguishing grammatical relations influenced by morphology. Our experiments, conducted with both encoder-only and decoder-only models, confirm that explicit morphological information contributes to more accurate syntactic analysis.

Keywords Korean dependency parsing Universal Dependencies (UD) UniMorph UniDive Morphosyntactic features Korean GSD treebank UDPipe LLM-based parsing (instruction tuning) Agglutinative languages

Building Korean UniDive

UniDive represents lexical words as dependency nodes while functional morphemes contribute features instead of separate syntactic units. The Korean UniDive dataset is built from the UD Korean GSD treebank:

  1. Treebank refinement. POS and morphological analyses in Korean GSD are corrected with POS tagging and named-entity recognition plus expert verification, following Sejong guidelines; dependency structures are not changed. The most frequent UPOS corrections are ADV → NOUN (3,607 tokens), NOUN → PROPN (2,654), and VERB → ADJ (1,065); the most frequent XPOS correction is XR+XSA+ETM → NNG+XSA+ETM (321).
  2. UniMorph-derived features. Each word is analyzed at the morpheme level and receives features by deterministic rules over morphemes and POS tags: Aspect, Case, Evident, Mood, NumType, Number, Person, Person[psor], Polite, PronType, Tense, VerbForm, and Voice (e.g., 이/가 → Case=Nom, 에게 → Case=Dat). Compared with K-UniMorph, new values include comitative, vocative, comparative, and allative case, hearsay and non-firsthand evidentiality, and causative-passive, reciprocal, and reflexive voice.
  3. Parsing experiments. An encoder-based parser (UDPipe 1) and a decoder-only LLM (Bllossom-3B, a Korean-adapted Llama 3.2 3B, instruction-tuned on the treebank) are trained on the standard GSD splits, each with and without the features.

Instruction-tuning example (Figure 1). The LLM receives the instruction and the CoNLL-U-style input and generates the head and relation columns:

[instruction]: 아래의 문장을 의존구조문법에 맞게 분석해줘
('Please parse the following sentence as a formal dependency parsing')
[output]:
1 학교     학교      NOUN  NNG      _          5  nsubj
2 분위기나 분위기+나 NOUN  NNG+JC   Case=Disj  1  flat
3 경관이   경관+이   NOUN  NNG+JKS  Case=Nom   1  conj
4 굉장히   굉장히    ADV   MAG      _          5  advmod
5 좋다     좋+다     ADJ   VA+EF    Mood=Ind   0  root
6 .        .         PUNCT SF       _          5  punct
('The school atmosphere and scenery are very good.')

Main Results

  • Large gain for the encoder-based parser. With morphosyntactic features, UDPipe 1 improves from 61.05 to 71.41 UAS and from 50.24 to 64.33 LAS (+14.09 LAS).
  • Smaller but consistent gain for the LLM parser. Instruction-tuned Bllossom-3B improves from 88.30 to 89.16 UAS and from 84.37 to 86.97 LAS (+2.6 LAS); the authors attribute the smaller gain to morphosyntax the LLM already captures in pretraining.
  • Morphology disambiguates relations. Explicit case features reduce subject–object confusion in non-canonical word orders, and verbal features (tense, aspect, modality) help analyze complex predicates with auxiliaries and honorifics.
  • Beyond Korean. Preliminary experiments on the UD Turkish Penn Treebank also show improved parsing accuracy with morphosyntactic features. The experiments use gold features; predicting them in context is left for future work.
Table 3. Korean dependency parsing on the Korean UniDive data (standard Korean GSD splits) with and without UniMorph-derived morphosyntactic (MS) features. UAS / LAS: unlabeled / labeled attachment score.
ModelSettingUASLAS
UDPipe 1Without MS features61.0550.24
With MS features71.4164.33
Bllossom-3BWithout MS features88.3084.37
With MS features89.1686.97

BibTeX

@misc{park2025enhancingkoreandependencyparsing,
      title={Enhancing Korean Dependency Parsing with Morphosyntactic Features},
      author={Jungyeul Park and Yige Chen and Kyuwon Kim and KyungTae Lim and Chulwoo Park},
      year={2025},
      eprint={2503.21029},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2503.21029},
}
More projects from KAIST MLP Lab →